Noise reduction and thinning method, system and equipment for cloud data of power distribution network, and medium

Abstract

The invention discloses a power distribution network point cloud data noise reduction and thinning method, system and device and a medium, and belongs to the technical field of power system point cloud data processing, and the method comprises the steps: constructing a tower structure model and a lead structure model containing structural feature parameters, extracting the geometric features of each point, calculating the similarity score, and achieving the point cloud classification; based on a normal vector, neighborhood geometry and a scoring result, structure confidence scoring is executed, and low confidence points and background points are removed; adopting a voxel division strategy driven by a structure trend to carry out self-adaptive thinning on the residual point clouds; and constructing a structural consistency loss function including main axis offset, curvature residual error and feature point space error, and executing fusion registration processing. According to the method, point cloud classification is guided through structure modeling, and the structure recognition precision is improved; the noise elimination accuracy is improved by combining confidence score; adopting structure self-adaptive thinning to retain key geometric features; structural consistency constraint optimization fusion registration is introduced, and the cloud processing effect and modeling quality of the power distribution network point are enhanced.

Description

Technical Field

[0001] This invention relates to the field of power system point cloud data processing technology, specifically to a method, system, equipment, and medium for noise reduction and thinning of distribution network point cloud data. Background Technology

[0002] With the development of UAV laser scanning and mobile measurement technologies, 3D point clouds, as an important data form for representing spatial information of real-world scenes, have been widely used in tasks such as power facility inspection and the construction of digital twin power grids. Especially in distribution network scenarios, acquiring high-density point cloud data, including towers, conductors, and their surrounding structures, has become one of the key data sources for smart grid information management.

[0003] Three-dimensional point cloud data collected for large-scale power distribution network scenarios often suffer from problems such as uneven density, high background noise, and weak structural features. Therefore, effective noise reduction, thinning, and registration processing must be performed before downstream structure identification, modeling, and analysis.

[0004] Current point cloud denoising and thinning methods mainly include density-based clustering methods (such as DBSCAN), spatial feature extraction algorithms (such as RANSAC), and uniform voxel grid filtering and statistical outlier removal. These methods are effective in general point cloud data processing tasks, but their generalization ability is weak in complex distribution network environments with obvious linear characteristics.

[0005] For example, existing methods often fail to accurately identify the continuity features of conductors or the vertical segmentation features of towers, which can easily lead to misclassification and structural damage. In addition, traditional point cloud registration methods (such as ICP) are mainly based on minimizing the matching of geometric point pairs, lacking structural semantic constraints, and making it difficult to achieve structural consistency alignment of targets such as tower main axes and conductor paths. Summary of the Invention

[0006] In view of the above-mentioned problems, the present invention is proposed.

[0007] Therefore, the technical problem solved by this invention is: how to effectively construct a structural model for 3D point cloud data containing towers, conductors, and background interference in complex power distribution network scenarios, accurately classify, denoise, and thin out the point cloud with structural preservation, and simultaneously achieve fusion processing with minimum registration accuracy residuals among multiple structural types, thereby improving the quality and efficiency of point cloud data in subsequent structural identification and modeling tasks.

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a method for noise reduction and thinning of point cloud data in a power distribution network, comprising,

[0009] Construct a tower structure model and a conductor structure model; the tower structure model includes a height range, vertical direction features, and a local normal concentration threshold; the conductor structure model includes a curvature range, linear extension direction, and a continuous point spacing threshold.

[0010] For each point in the 3D point cloud data of the power distribution network, the spatial height, normal vector direction, neighborhood point arrangement direction, and local curvature are extracted. Similarity calculations are performed with the tower structure model and the conductor structure model respectively. Based on the similarity score results, the point cloud is classified into tower point cloud, conductor point cloud, and background point cloud.

[0011] For each point in the classification results, structural confidence scoring is performed. The input features include normal vector, neighborhood geometry, and similarity score. The output is a structural score value.

[0012] Delete tower point cloud points and conductor point cloud points whose score values ​​are lower than the preset threshold, and delete all background point cloud points;

[0013] Set the voxel mesh size and thinning weight for the tower point cloud and conductor point cloud after the deletion operation, and perform structure adaptive thinning processing.

[0014] The point cloud of the tower and the point cloud of the conductor after thinning are fused and registered to minimize the structural consistency loss function. The structural consistency loss function includes the tower principal axis offset term, the conductor curve residual term, and the spatial offset term of the structural feature points. The fusion and registration output is used as the final point cloud data.

[0015] As a preferred embodiment of the noise reduction and thinning method for point cloud data of power distribution network described in this invention, the construction of the tower structure model and conductor structure model includes: extracting the spatial height features of each point in the point cloud, calculating the average height value of the local area, filtering points belonging to a set height range, and forming a first candidate point set;

[0016] Vertical features are extracted from the first candidate point set. By calculating the angular relationship between the main direction of the point and the vertical direction, points that meet the vertical feature requirements are identified, forming the second candidate point set.

[0017] Perform normal vector statistics on the neighborhood of each point in the second candidate point set, calculate the normal vector consistency index, and select points that meet the local normal concentration threshold to form the tower structure point set;

[0018] Based on the statistical characteristics of the point set of the tower structure, a tower structure model is defined that includes height range, vertical direction characteristics, and local normal concentration threshold.

[0019] As a preferred embodiment of the noise reduction and thinning method for distribution network point cloud data described in this invention, the construction of the tower structure model and conductor structure model further includes identifying a continuous point sequence in the point cloud data where the distance between adjacent points is less than a set threshold, filtering point segments that meet the continuous point spacing threshold, and forming a conductor candidate point set.

[0020] For each segment of the candidate points in the conductor set, calculate the local curvature characteristics, filter out segments with curvature exceeding the set curvature range, and retain segments with curvature that meet the requirements.

[0021] Within the retained point segments, direction vector fitting is performed to extract the linear extension direction, identify point segments whose direction changes are within the allowable range, and form a set of conductor structure points;

[0022] A conductor structure model is constructed based on the curvature range, linear extension direction, and point spacing parameters of the conductor structure point set.

[0023] As a preferred embodiment of the noise reduction and thinning method for distribution network point cloud data described in this invention, the step of classifying the point cloud into tower point cloud, conductor point cloud, and background point cloud based on similarity scoring results includes: sequentially setting a weighted priority factor in the tower structure model scoring based on the distance between the spatial height of each point and the estimated center of the tower's main axis; calculating the difference between the spatial height, normal vector direction, and neighborhood normal concentration of each point and the height range, vertical direction feature, and neighborhood normal concentration threshold set in the tower structure model, and combining the weighted priority factor to form the tower structure model scoring value;

[0024] The scoring value of the traverse structure model is formed by combining the directional deviation between the neighborhood point arrangement direction of each point and the linear extension direction set in the traverse structure model, the difference between the local curvature and the curvature range set in the traverse structure model, and the deviation between the distance between the point and the adjacent point and the continuous point spacing threshold set in the traverse structure model.

[0025] Extract the mean scores of the tower structure model and the mean scores of the conductor structure model from the points within the neighborhood of each point. Then, perform weighted correction on the tower structure model score and the conductor structure model score for each point to obtain the corrected tower structure model score and the corrected conductor structure model score.

[0026] When the score of the corrected tower structure model is greater than the score of the corrected conductor structure model, and the score difference between the two is greater than the score difference threshold, and the score of the corrected tower structure model is greater than the classification confidence threshold, it is determined to be a tower point cloud category.

[0027] When the score of the corrected conductor structure model is greater than the score of the corrected tower structure model, and the score difference between the two is greater than the score difference threshold, and the score of the corrected conductor structure model is greater than the classification confidence threshold, it is determined to be a conductor point cloud category.

[0028] If the point cloud category for towers and the point cloud category for conductors are not met, the point cloud category is determined to be the background point cloud category.

[0029] As a preferred embodiment of the noise reduction and thinning method for distribution network point cloud data described in this invention, the input features include normal vector, neighborhood geometry, and similarity score. The output structure score value includes extracting the normal vector direction change features of each point, the local distribution geometry features of points within the neighborhood, and the similarity strength features corresponding to the structure model score value of the category, and constructing a structure confidence feature vector.

[0030] The magnitude of change in normal vector direction, the compactness index of neighborhood point space, and the similarity score intensity index are converted into confidence sub-feature values, respectively.

[0031] The structural confidence scoring function is constructed by combining the confidence sub-features according to the preset weighted fusion rules.

[0032] Input the structural confidence feature vector into the structural confidence scoring function to generate a structural score value for each point;

[0033] When the structural score of a point cloud point of a tower or a point cloud point of a conductor is lower than the structural score threshold, the point is deleted from the corresponding point cloud set.

[0034] Delete all points in the background point cloud set.

[0035] As a preferred embodiment of the noise reduction and thinning method for power distribution network point cloud data described in this invention, the following steps are included: performing adaptive structural thinning processing, constructing a vertical structural trend estimation curve based on the gradient distribution of the change amplitude of the normal vector direction of each point in the tower point cloud in the vertical direction; dividing the tower point cloud into multiple structural segments according to the curve; setting a constant height interval as the voxel partitioning unit size within each structural segment, and extracting representative points covering the key nodes of the structural trend curve to form a representative tower point set;

[0036] Based on the arrangement direction of each point in the traverse point cloud, a fitting path for the traverse principal axis is generated; the distance offset value from each point to the fitting path is calculated; the fitting path is used as a reference axis to divide the projection segments with fixed spacing, and points with consistent arrangement direction and minimum local curvature are extracted in each segment to form a representative point set of the traverse.

[0037] Edge continuity analysis is performed on the representative point sets of towers and conductors respectively, and boundary points with abrupt changes in location distribution and redundant points with local density anomalies exceeding the preset redundancy threshold are removed.

[0038] As a preferred embodiment of the noise reduction and thinning method for distribution network point cloud data described in this invention, the step of performing fusion registration processing on the thinned tower point cloud and conductor point cloud, and minimizing the structural consistency loss function includes,

[0039] Extract a set of structural feature points from the tower point cloud; fit the tower principal axis vector in the vertical direction based on the set of structural feature points, and calibrate the coordinates of the tower principal axis starting point;

[0040] Extract a continuously arranged set of points from the traverse point cloud; perform spatial fitting based on this set to generate a traverse curvature trend function;

[0041] Based on the temporal or spatial overlap of point cloud acquisition, the structural feature point correspondence between tower point cloud and conductor point cloud is calibrated.

[0042] A structural consistency loss function is constructed, which includes: a tower principal axis offset term, representing the angular deviation between the tower point cloud principal axis vector and the reference principal axis vector, as well as the Euclidean distance residual between the principal axis starting points; a traverse curve residual term, representing the integral value of the curvature offset residual between the traverse point cloud curvature trend function and the reference curvature function in the corresponding segment; and a structural feature point spatial offset term, representing the sum of squared three-dimensional spatial coordinate residuals between corresponding structural feature points.

[0043] The optimal spatial transformation matrix is ​​calculated by minimizing the structural consistency loss function. The spatial transformation matrix is ​​then applied to perform a fusion registration operation on the tower point cloud and the conductor point cloud. The fused and registered point cloud data is then output as the final point cloud data result.

[0044] This invention provides a noise reduction and thinning system for point cloud data of power distribution networks.

[0045] To solve the above technical problems, the present invention provides the following technical solution: a noise reduction and thinning system for point cloud data of power distribution network, comprising: a construction module configured to construct tower structure models and conductor structure models;

[0046] The feature extraction and classification module is configured to extract the spatial height, normal vector direction, neighborhood point arrangement direction and local curvature of each point in the three-dimensional point cloud data of the power distribution network, and perform similarity calculations with the tower structure model and conductor structure model respectively, and classify the point cloud into tower point cloud, conductor point cloud and background point cloud based on the similarity score results.

[0047] The structure scoring module is configured to perform structure confidence scoring on each point in the classification results, extract the normal vector direction, neighborhood geometry and similarity score, and generate a structure score value.

[0048] The removal module is configured to delete tower point cloud points and conductor point cloud points whose structural score values ​​are lower than a preset threshold, and to delete all background point cloud points;

[0049] The thinning module is configured to set the voxel mesh size and thinning strategy for the removed tower point cloud and conductor point cloud respectively, and perform structure-adaptive thinning processing.

[0050] The registration module is configured to perform a fusion registration operation on the thinned tower point cloud and the conductor point cloud, minimizing the structural consistency loss function.

[0051] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method for noise reduction and thinning of point cloud data in a power distribution network.

[0052] The present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the steps of the method for noise reduction and thinning of point cloud data in a power distribution network.

[0053] The beneficial effects of this invention are as follows: By constructing a priori model that includes the structural features of towers and conductors, this invention can accurately identify target areas with structural differences in the power distribution network during point cloud classification, thus improving the accuracy of point cloud classification. A structural confidence scoring function is constructed by combining spatial height, normal variation, neighborhood arrangement, and model scoring information to quantitatively judge point-level structural consistency, improving the accuracy of identifying and removing noise points and misjudged points. An adaptive thinning method based on structural trend estimation is adopted, and a voxel partitioning strategy is reasonably set in different structural segments, achieving point cloud simplification while maintaining local structural integrity. In the fusion registration stage, a structural consistency loss function containing tower principal axis offset, conductor curvature residual, and spatial error of structural feature points is constructed to perform registration optimization on the thinned point cloud, improving the structural accuracy and stability of point cloud alignment. This invention is suitable for point cloud preprocessing and structural modeling tasks in complex power distribution network scenarios. Attached Figure Description

[0054] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 This is a flowchart illustrating a method for noise reduction and thinning of point cloud data in a power distribution network, as provided in one embodiment of the present invention. Detailed Implementation

[0056] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0057] Example 1, referring to Figure 1 This is one embodiment of the present invention, which provides a method for noise reduction and thinning of distribution network point cloud data, including:

[0058] Step S100: Construct a tower structure model and a conductor structure model; the tower structure model includes a height range, vertical direction features, and a local normal concentration threshold; the conductor structure model includes a curvature range, linear extension direction, and a continuous point spacing threshold.

[0059] Step S200: For each point in the 3D point cloud data of the power distribution network, extract the spatial height, normal vector direction, neighborhood point arrangement direction, and local curvature, and perform similarity calculations with the tower structure model and conductor structure model respectively. Based on the similarity score results, classify the point cloud into tower point cloud, conductor point cloud, and background point cloud.

[0060] Step S300: Perform structural confidence scoring on each point in the classification results. The input features include normal vector, neighborhood geometry, and similarity score. The output is the structural score value.

[0061] Step S400: Delete tower point cloud points and conductor point cloud points with score values ​​lower than the preset threshold, and delete all background point cloud points;

[0062] Step S500: Set the voxel mesh size and thinning weight for the tower point cloud and conductor point cloud after the deletion operation, and perform adaptive structural thinning processing;

[0063] Step S600: Perform fusion registration processing on the thinned tower point cloud and the conductor point cloud, and minimize the structural consistency loss function. The minimized structural consistency loss function includes the tower principal axis offset term, the conductor curve residual term, and the structural feature point spatial offset term. The fusion registration output result is used as the final point cloud data.

[0064] In distribution network point cloud data, tower structures and conductor structures often exhibit significant differences in geometric characteristics. Towers are mainly distributed vertically with large height spans and concentrated local normals, while conductors show continuous horizontal extension with small local curvature and strong directional consistency. However, in actual data acquisition, environmental occlusion, equipment vibration, and uneven point cloud density lead to blurred boundaries between structures and a large number of data noise and redundant points, making it difficult to distinguish different structures using traditional point cloud processing methods, thus affecting subsequent analysis and modeling. This embodiment introduces a structure model-driven hierarchical processing method. First, it defines the geometric structural standards of towers and conductors, establishes structural model parameters, and then guides classification, scoring, and thinning steps through structure matching. This achieves structure-aware cleaning and simplification of distribution network point clouds, improving the accuracy and efficiency of structure identification.

[0065] The method described in this embodiment starts with structural modeling, establishing a point cloud structural model of towers and conductors as the basic constraint for the entire process. Multi-dimensional feature extraction and analysis are performed on each point in the point cloud using indicators such as spatial height, directional distribution, and curvature range, constructing a matching scoring mechanism between points and the model. Subsequently, the confidence level of the point classification results is adjusted by combining the scoring values ​​and local structural statistical results, eliminating points that may have structural interference or density anomalies to avoid the propagation of local misjudgments. Based on the structural category, the point cloud is divided into different thinning regions, and a thinning method matching the structural trend is adopted to reduce data redundancy. Finally, in the fusion stage, structural consistency constraints are introduced to establish a registration residual function. Through multi-constraint optimization of the tower principal axis, conductor curvature, and the correspondence of feature points, the point cloud results maintain accurate representation of the overall geometric relationship while ensuring coordinate system consistency, meeting the quality requirements for direct use in subsequent digital modeling or 3D reconstruction.

[0066] Reference Figure 1The processing flow shown first involves preliminary screening of the original power distribution network point cloud data to extract point sets with structural characteristics. Tower and conductor structure models are then constructed to guide subsequent classification. During the processing of each point, features such as spatial height, normal vector direction, neighborhood arrangement direction, and local curvature are extracted sequentially. These features are compared item by item with the parameter ranges of the two structural models, and a score is calculated. This score is then corrected using the local neighborhood statistical average to form the classification criteria. In the structural scoring stage, a confidence score is constructed using three indicators: normal consistency, geometric compactness, and structural matching strength. This score serves as the standard for retention and rejection, ensuring the retention of points with high structural stability. For the classified point cloud, a thinning strategy is applied based on its category, performing voxel thinning and structural trend thinning respectively. Finally, the thinned tower and conductor point clouds are input into the structural consistency registration module. Registration calculations are performed using three residual functions: tower principal axis, conductor curvature, and structural point coordinate offset. This completes the final structural alignment and outputs a simplified point cloud result with high consistency, low noise, and complete features.

[0067] Example 2, an embodiment of the present invention, provides a method for noise reduction and thinning of distribution network point cloud data based on the previous embodiment, including:

[0068] Specifically, step S100 constructs a tower structure model and a conductor structure model; the tower structure model includes a height range, vertical direction features, and a local normal concentration threshold; the conductor structure model includes a curvature range, linear extension direction, and a continuous point spacing threshold. The construction of the tower structure model includes the following steps A1-A4:

[0069] A1: Extract the spatial height features of each point in the point cloud, calculate the average height value of the local area, filter points that belong to the set height range, and form the first candidate point set.

[0070] Specifically, for any point in the distribution network point cloud data, a spherical neighborhood is established centered on that point. Within a set radius (e.g., 1.0 meter), the spatial Z-coordinate values ​​of all points within the neighborhood are statistically analyzed, and their average value is calculated as the local average height of that point. The difference between the spatial Z-value of that point and the average height of its neighborhood is compared. When the difference is within a preset height range (e.g., 3 meters to 40 meters), the point is considered to likely belong to the tower structure area and is included in the first candidate point set.

[0071] A2: Extract vertical features from the first candidate point set, and identify points that meet the vertical feature requirements by calculating the angular relationship between the main direction of the point and the vertical direction, thus forming the second candidate point set.

[0072] Specifically, for each point in the first candidate point set, a covariance matrix is ​​constructed using its neighborhood point set, and principal component analysis (PCA) is performed to obtain the local principal direction vector of that point. Using the vertical unit vector (0,0,1) as a reference, the angle between the principal direction vector and the vertical direction is calculated. When this angle is less than a set threshold (e.g., 15°), the point is determined to have obvious vertical structural features and is included in the second candidate point set.

[0073] A3: Perform normal vector statistics on the neighborhood of each point in the second candidate point set, calculate the normal vector consistency index, and select points that meet the local normal concentration threshold to form the tower structure point set.

[0074] Specifically, firstly, the normal vector of each point in the second candidate point set is calculated. Then, the set of all normal vectors within its neighborhood is obtained. By calculating the average cosine of the angle between the normal vector of this point and the normal vectors of each point in the neighborhood, the normal vector consistency index of this point is obtained. For example, if the average cosine of the normal vectors in the neighborhood is greater than 0.95 (indicating that the normal vectors tend to be concentrated), the point is considered to have good normal vector consistency and is retained in the tower structure point set. This index can effectively identify point cloud distribution characteristics with obvious columnar or tower-like structures.

[0075] A4: Based on the statistical characteristics of the tower structure point set, define a tower structure model that includes height range, vertical direction characteristics, and local normal concentration threshold.

[0076] After constructing the point set of the tower structure, the maximum and minimum values ​​of its height range are statistically analyzed, and the mean and standard deviation of the principal direction deviation, as well as the statistical mean of the normal consistency index, are recorded to define the parameters of the tower structure model. Specifically, the height range is used to limit the vertical distribution of the tower point cloud, the vertical direction features are used to constrain the principal direction of the point cloud structure, and the local normal concentration threshold is used to identify the degree of structural regularity. The completed structural model serves as a matching reference for subsequent point cloud classification and scoring processing.

[0077] The construction of the conductor structure model includes the following steps B1-B4:

[0078] B1: Identify a sequence of consecutive points in the point cloud data where the distance between adjacent points is less than a set threshold, filter out point segments that meet the consecutive point spacing threshold, and form a candidate point set for the traverse.

[0079] Specifically, the original point cloud data is first spatially sorted, and a sequence of point pairs with local continuity is extracted by coordinates or spatial index. The Euclidean distance between adjacent points is calculated, and when the distance between multiple consecutive points is less than a set threshold (e.g., 0.5 meters), the continuous point sequence is marked as a candidate segment. The set of all point segments that satisfy this distance constraint constitutes the candidate point set for the traverse. Such continuous and densely arranged point segments are usually the main compositional pattern of aerial traverses.

[0080] B2: Calculate the local curvature characteristics of each point sequence in the candidate point set of the conductor, filter out point segments whose curvature exceeds the set curvature range, and retain point segments whose curvature meets the requirements.

[0081] In this step, a sliding window method or a three-point fitting method is used to estimate the local curve of the candidate segments. Specifically, a local spline curve or quadratic curve is used to fit each point sequence, and the derivative of the fitted curve is calculated at each point to determine the curvature value. When the mean or maximum curvature value of the overall curvature of the segment is within a preset curvature range (e.g., 0–0.3 / m), it is considered to have a regular conductor shape and stable curvature change, and the segment is retained; otherwise, it is considered to be abnormally twisted or a non-conductor structure and is discarded.

[0082] B3: In the retained point segments, perform direction vector fitting processing, extract the linear extension direction, identify point segments with the degree of direction change within the allowable range, and form a set of conductor structure points.

[0083] Specifically, the process involves: constructing the covariance matrix of the point cloud for each segment and extracting the principal direction vector as the linear extension direction of that segment; then, calculating the angle between the principal direction vectors of adjacent segment segments. If this angle is less than a set threshold (e.g., 10°), it indicates that the extension direction remains stable, and the segment is merged with the previous segment into a linear continuous region; if the direction changes significantly, the segment is discarded. Finally, all segment segments that meet the linear extension direction requirement are merged to form the traverse structure point set.

[0084] B4: Construct a conductor structure model based on the curvature range, linear extension direction, and point spacing parameters of the conductor structure point set.

[0085] Finally, the geometric constraint parameters of the traverse structure are extracted by statistically analyzing the average curvature, principal direction vector, directional fluctuation range, and continuous point spacing distribution of each segment in the traverse structure point set. The curvature range defines the bending scale of the traverse, the linear extension direction characterizes its spatial extension trend, and the point spacing parameter reflects the density and distribution pattern of the traverse. These three parameters together constitute the traverse structure model, which serves as the structural prior standard for the traverse point cloud in subsequent point cloud classification and structure scoring processing.

[0086] Further, step S200 extracts the spatial height, normal vector direction, neighborhood point arrangement direction, and local curvature of each point in the 3D point cloud data of the distribution network, and performs similarity calculations with the tower structure model and conductor structure model respectively. Based on the similarity score results, the point cloud is classified into tower point cloud, conductor point cloud, and background point cloud, including the following steps C1-C6:

[0087] C1: Based on the distance between the spatial height of each point and the estimated center of the tower's main axis, set the weighting priority factor in the tower structure model score; calculate the difference between the spatial height, normal vector direction, and neighborhood normal concentration of each point and the height range, vertical direction characteristics, and neighborhood normal concentration threshold set in the tower structure model, and combine them with the weighting priority factor to form the tower structure model score.

[0088] C2: Combine the directional deviation between the neighborhood point arrangement direction of each point and the linear extension direction set in the traverse structure model, the difference between the local curvature and the curvature range set in the traverse structure model, and the deviation between the distance to adjacent points and the continuous point spacing threshold set in the traverse structure model to form the traverse structure model score.

[0089] C3: Extract the mean scores of the tower structure model and the mean scores of the conductor structure model from the points within the neighborhood of each point. Then, perform weighted correction on the tower structure model score and the conductor structure model score for each point to obtain the corrected tower structure model score and the corrected conductor structure model score.

[0090] C4: When the score of the corrected tower structure model is greater than the score of the corrected conductor structure model, and the score difference between the two is greater than the score difference threshold, and the score of the corrected tower structure model is greater than the classification confidence threshold, it is determined to be a tower point cloud category.

[0091] C5: When the score of the corrected conductor structure model is greater than the score of the corrected tower structure model, and the score difference between the two is greater than the score difference threshold, and the score of the corrected conductor structure model is greater than the classification confidence threshold, it is determined to be a conductor point cloud category.

[0092] C6: If the tower point cloud category and the conductor point cloud category are not met, it is determined to be the background point cloud category.

[0093] Specifically, in step C1, the tower structure model score is formed by combining the weighted priority factors, including the following steps C11-C14:

[0094] Step C11: Construct the normal concentration index K p For the point p to be processed, a neighborhood is established in its three-dimensional space with a fixed search radius r = 0.3m. The unit normal vector n for all points in the neighborhood i Constructing the structure tensor C p :

[0095]

[0096] Computing the structure tensor C pLet the eigenvalues ​​of point p be λ1 and λ2, respectively, and define the normal concentration index of point p as:

[0097]

[0098] Where ε = 10 -6 A constant used to suppress division-by-zero errors. This index characterizes the concentration of the neighborhood normal vectors along the principal direction, K. p The larger the value, the more concentrated the normal direction and the higher the structural regularity.

[0099] Step C12: Set the weighted priority factor ω z The height coordinate z of point p p Combined with the tower's main axis, estimate the height coordinate z of the center. c Calculate the vertical distance:

[0100] d z =|z p -z c |

[0101] Based on this distance, define a height-weighted priority factor:

[0102]

[0103] Where σ is a height scale parameter, preferably set to half the width of the height range of the tower structure model. For example, when the average tower height is H = 18m, σ = 9m is taken.

[0104] Step C13: Calculate the structural difference index. For point p, calculate the difference between the point and the tower structure model for the three key structural features: height difference δ. h Let z be the center of the height interval in the tower structure model. mid The half-width is z range Then we have:

[0105]

[0106] Normal difference δ n Let the unit normal vector of point p be n. p Let the unit vector in the vertical direction be e. z =(0,0,1) T Then the cosine of the included angle is:

[0107] δ n =1-|n p ·e z |

[0108] Concentration Difference δ f Let the concentration threshold of the tower structure model be κ0, which is derived from all κ points in the model point set.p The 70th percentile of the value is:

[0109]

[0110] After normalization, this difference reflects the degree of normal convergence of points and the degree of difference in the model.

[0111] Step C14: Calculate the integrated score. Combine the three difference indicators mentioned above, use a weighted method to integrate them, and combine them with the priority factor to obtain the final score value.

[0112] S tower =ω z ·(α·(1-δ h )+β·(1-δ n )+γ·(1-δ f ))

[0113] Where α, β, and γ represent weighting coefficients, respectively.

[0114] Final score S tower ∈[0,1], the higher the value, the better the point matches the tower structure model in terms of shape.

[0115] It should be noted that in step C14, the fitting of the weighting coefficients includes the following steps C141-C144:

[0116] Step C141: Segmentation of the tower structure. For multiple manually labeled tower point cloud samples, segment the structure along the height axis (Z-axis), defining three typical structural segments: Lower segment (base): closer to the ground, robust structure; Middle segment (trunk): the upright section of the tower, constituting the main body of the structure; Upper segment (top / crossarm): more transverse structures, greater morphological variation. Let r1, r2, and r3 represent the relative proportions of these three segments in the overall tower structure, satisfying r1 + r2 + r3 = 1, such as r1 = 0.25, r2 = 0.50, r3 = 0.25.

[0117] Step C142: Statistical analysis of three feature differences. Within each structural segment, calculate the average difference of the labeled sample points for the following three structural features: mean height difference. Normalized difference between current point height and paragraph center height; mean of normal angle The cosine distance of the angle between the current point's normal and the vertical direction; normal concentration deviation. The difference between the normal concentration of the current point's neighborhood and the threshold κ0, where i = 1, 2, 3 represent the corresponding structural segment numbers, is normalized by taking the reciprocal of the mean of the above differences to obtain the structural dominance index within the segment:

[0118]

[0119] in, This is used to ensure the normalization of the two dominant weights, i.e.

[0120]

[0121] Step C143: Generate the final weighting coefficients by weighting each paragraph's dominant indicator with its spatial proportion.

[0122]

[0123] Since all w values ​​have been normalized, the fusion coefficient also satisfies α+β+γ=1.

[0124] Step C144: Sample validation and weight optimization. Apply the calculated (α,β,γ) to the actual point cloud classification task on the sample set, record the recognition accuracy, false positive rate and false negative rate of the tower points in each structural segment, and comprehensively form a performance evaluation index (such as F1 score).

[0125] If the performance index does not reach the set threshold (e.g., F1-score ≥ 0.9), the dominance coefficients of each segment are fine-tuned, and the fusion weights are further fitted by gradient descent or minimum mean square error optimization until the target performance is met. The finally determined α, β, γ are used as the optimal fusion weight coefficients in the tower structure model scoring function of this invention for actual point cloud analysis.

[0126] Furthermore, in step C2, the calculation of the conductor structure model score includes the following steps C21-C24:

[0127] Step C21: Calculate the direction deviation term and obtain the set of points in the neighborhood of point p. Principal component analysis (PCA) is used to fit the principal direction vector u. p , as the local linear arrangement direction of point p;

[0128] This direction vector is compared with the preset linear extension direction u in the conductor structure model. ref The included angle is calculated, and the direction deviation term is defined as follows:

[0129] δ dir =1-|u p ·u ref |

[0130] Here, · represents the vector dot product operation, and the result is processed by absolute value to ensure sensitivity to direction consistency.

[0131] Step C22: Calculate the curvature deviation term by performing a local fitting (e.g., fitting a quadratic curve) to calculate the local curvature value of point p. Compare it with the reference curvature value set in the wire structure model Perform difference normalization and define the curvature deviation term as follows:

[0132]

[0133] Step C23: Calculate the point spacing deviation term. The Euclidean distance between point p and its nearest neighbor q along the principal direction is denoted as d. pq Let the predefined reference value for the continuous point spacing in the conductor structure model be d. rcf Then the point spacing deviation term is defined as:

[0134] Where, d max This represents the maximum allowable deviation threshold for the distance between dots.

[0135] Step C24: Calculate the structural score of the conductor structure. Weight and fuse the three deviation terms mentioned above to obtain the structural score S of point p relative to the conductor structure model. wire The calculation formula is as follows:

[0136] S wire =θ1·(1-δ dir )+θ2·(1-δ curv )+θ3·(1-δ dist )

[0137] Where θ1,θ2,θ3∈[0,1] are the weight coefficients of each term, satisfying:

[0138] θ1+θ2+θ3=1

[0139] The aforementioned weights can be optimized by fitting typical traverse sample point cloud data through a training method that minimizes structural consistency residuals, so as to ensure the strongest consistency between the model scoring results and the manually labeled results.

[0140] It should be noted that in actual power distribution network scenarios, the conductor point cloud structure has the following engineering characteristics: the conductor is a tension-guided structure, and its arrangement direction is highly consistent over a large range; it is slightly disturbed by wind load, electric arc, etc., but it needs to maintain continuity to maintain power supply reliability; the non-uniform scanning of laser sampling causes the point spacing of the same conductor segment to be variable, but it does not affect the judgment of structural continuity.

[0141] Therefore, when evaluating the scoring intensity of the three features, directional consistency should be the primary criterion, followed by curvature continuity, and finally, point spacing uniformity. The training of the weight coefficients θ1, θ2, and θ3 includes the following steps C241-C244:

[0142] Step C241: Establish a structural sample set and collect high-density point cloud samples from several typical conductor sections. Each segment P j Contains N j Each point ensures coverage of typical shapes such as straight lines, turns, and downward segments.

[0143] Step C242: Calculate the standard deviation of the structural bias. In each segment, calculate the sample standard deviation of the following three types of structural bias indicators for each point, defined as: σ dir,j : Standard deviation of orientation (reflecting the degree of fluctuation in the orientation); σ curv,j : Standard deviation of local curvature deviation (reflecting the strength of curvature disturbance); σ dist,j Standard deviation of point spacing (reflects the degree of density non-uniformity).

[0144] Step C243: Construct a structural reliability inverse weighting, and calculate the mean of all paragraphs based on the overall uncertainty level of each type of deviation in the sample set:

[0145]

[0146] Based on the inverse principle of reliability (i.e., the smaller the standard deviation, the higher the stability, and the greater the weight), the initial weight values ​​are set as follows:

[0147]

[0148] Step C244: Normalize the above weights to obtain a weight allocation that conforms to the weighting rules:

[0149]

[0150] Furthermore, in step C3, the mean neighborhood score is calculated. A neighborhood of fixed radius is constructed in three-dimensional space, centered on the current processing point p, denoted as: Where, x p =(x p ,y p ,z p () represents the three-dimensional coordinates of point p, r c This represents the neighborhood search radius, which is preferably set to r. c =0.5m, q represents any point in the neighborhood, and ||·|| represents the Euclidean distance norm.

[0151] Calculate the mean of the original structure score values ​​for all points within the neighborhood of point p:

[0152]

[0153] in, The score value of the tower structure model at point q; The score of the traverse structure model at point q; Indicates the number of points in the neighborhood; Let p represent the average score within the neighborhood of point p.

[0154] The original score of point p is weighted and fused with the mean of its neighborhood to obtain the corrected score:

[0155]

[0156] Weight, This indicates the corrected tower structure score; Indicates the corrected conductor structure score; η T The original retention factor (range [0,1]) represents the tower structure score; η W The original retention factor (range [0,1]) represents the conductor structure score.

[0157] To achieve a weighted fusion strategy with engineering usability and adaptability, this invention preferably uses a combination of empirical evaluation and local score stability analysis to calculate the weighting factors: Sample group evaluation: Select tower point cloud and conductor point cloud samples respectively, and perform statistical analysis based on the structural score fluctuation.

[0158] Rating variance measurement: For each point p in the point cloud sample, the local variance of the structure rating is defined as:

[0159]

[0160] Among them, S q Assign a score to the neighboring points (towers or conductors). The mean score of the neighborhood is used as the basis for calculating the mean and variance of the two types of structured point cloud samples, respectively, to obtain their average score volatility.

[0161] The volatility index of the structural score is used as a weight for inverse mapping:

[0162]

[0163] Where k represents the regularization coefficient, which is used to adjust the degree of influence of score fluctuation on weight. It is preferred to set k = 10. When the score fluctuation is small (structure is regular), η→1, and the original score is retained. When the fluctuation is large (structure is unstable), η→0, and the neighborhood correction is enhanced.

[0164] After training and validation using measured point cloud data from multiple power distribution network scenarios, this method has been shown to obtain an average η value in tower structures. T ≈0.83, η is obtained in the wire structure W The value is approximately 0.61, which is superior to manually set weights in terms of both classification accuracy and dilution stability.

[0165] Rating difference threshold Δt h is used to constrain the minimum structural difference between the two types of scores, in order to avoid misclassification caused by ambiguous classification boundaries. This threshold is set through the following steps:

[0166] Data preparation: Select a set of point cloud samples with known category labels, including tower point clouds and conductor point clouds, to form the training dataset.

[0167] Score Acquisition: For each point p in the training set, calculate its corrected tower structure score. With wire structure scoring And calculate the score difference:

[0168]

[0169] Statistical analysis: ΔS for all training points p Construct a frequency distribution histogram and analyze the degree of concentration of differences;

[0170] Parameter extraction: Set the score difference threshold Δ t h represents the 75th percentile of this difference distribution, meaning that 75% of the training point score differences are greater than this value, thus ensuring that most classified points have clear structural distinction.

[0171]

[0172] Example explanation: If the median score difference in the training sample is 0.08 and the 75th percentile is 0.11, then select:

[0173] Δ th =0.11

[0174] Classification confidence threshold S th Points ultimately assigned to a specific structural category must have a score higher than a certain strength threshold to prevent misclassification of points with excessively low scores. The selection rules are as follows:

[0175] Group processing: Calculate the corrected structure score for both tower point clouds and conductor point clouds in the training set.

[0176]

[0177] Single-class distribution analysis: Construct cumulative distribution functions (CDF) for the two groups of scores respectively, and observe their central tendency;

[0178] Weigh the weakest confidence threshold: Select the 85th percentile of the two sets of scores, that is, the 85th percentile range of the most confident scores in each category.

[0179]

[0180] Unified threshold setting: The smaller of the two values ​​is taken as the general confidence threshold for classification, to ensure a consistent judgment standard across different categories.

[0181]

[0182] Example explanation: If the 85th percentile of the tower rating distribution is 0.71 and the conductor rating is 0.68, then the following settings are uniformly applied:

[0183] S th =0.68

[0184] Further, step S300 performs structural confidence scoring on each point in the classification result. The input features include normal vector, neighborhood geometry, and similarity score. The output structural score value is used in step S400 to delete tower point cloud points and conductor point cloud points with score values ​​lower than a preset threshold, and to delete all background point cloud points. This includes the following steps D1-D6:

[0185] D1: Extract the normal vector direction change features of each point, the local distribution geometric features of points within the neighborhood, and the similarity strength features corresponding to the structural model score of the category, and construct the structural confidence feature vector;

[0186] D2: Convert the normal vector direction change magnitude, neighborhood point spatial compactness index, and similarity score intensity index into confidence sub-feature values ​​respectively;

[0187] D3: Combine the sub-feature values ​​of each confidence level according to the preset weighted fusion rules to construct the structural confidence scoring function;

[0188] D4: Input the structural confidence feature vector into the structural confidence scoring function to generate the structural score value for each point;

[0189] D5: When the structural score of a point cloud point of a tower or a point cloud point of a conductor is lower than the structural score threshold, the point is deleted from the corresponding point cloud set.

[0190] D6: Delete all points in the background point cloud set.

[0191] Specifically, for each point q in the point cloud, three types of structural feature indicators are extracted to construct the structural confidence feature vector v. q :

[0192] Normal wave angle θ q The unit normal vector of point q is n. q The mean of the unit normal vectors of all points in its neighborhood is but:

[0193]

[0194] Where, nq Let represent the unit normal vector of point q; θ represents the mean of the neighborhood normal vectors of point q; q It represents the normal undulation angle, reflecting the directional stability of a point.

[0195] Neighborhood compactness index ρ q :

[0196]

[0197] Where, N q ρ represents the number of points in the neighborhood of point q (with a neighborhood radius of r); r represents the search radius, preferably set to 0.3m; q It represents the density of neighboring points within a unit volume, reflecting the degree of spatial aggregation.

[0198] Structural model similarity score (normalized) ξ q :

[0199]

[0200] Among them, s q This represents the structural model score for the current point; s min ,s max Indicates the minimum and maximum values ​​in the score; ξ q This represents the normalized similarity score, used to compare the degree of matching of structural features of different categories.

[0201] Finally, the three structural indicators are integrated into a feature vector:

[0202]

[0203] The three structural features are transformed into confidence sub-features in the interval [0,1] through a function mapping: normal consistency confidence C. θ (q):

[0204]

[0205] Where, σ θ The desired scale for the change in direction is preferably 15° (input in radians);

[0206] Compactness confidence

[0207]

[0208] in, This represents the maximum neighborhood density value observed among all points.

[0209] Structure matching confidence level C ξ (q):

[0210] C ξ (q)=ξ q

[0211] The three confidence features are fused in a linear weighted manner to construct a structural scoring function:

[0212]

[0213] Among them, C score (q) represents the structural confidence score for point q. Represents the weighting coefficients, satisfying:

[0214] v q Input the above scoring function to obtain the structural score C for each point. score (q)∈[0,1], the higher the score, the more reliable the point is in terms of local geometry and global structure.

[0215] Set the structural scoring threshold τ s ∈[0,1], for each point q classified as a tower point or a traverse point, if:

[0216] C score (q)<τ s

[0217] Then delete point q from the corresponding point cloud set. τ s This represents the structural scoring threshold.

[0218] In this embodiment, a structural scoring threshold τ is determined for removing low structural confidence points. s Experimental analysis was conducted on multiple typical power distribution network point cloud datasets. Using point clouds with manually labeled ground truth values ​​collected from actual engineering projects as benchmark data, structural scoring and removal experiments were performed at different scoring threshold levels (e.g., 0.3, 0.4, 0.5, 0.6, 0.7). The retention rate and false removal rate curves of the true structural points in the remaining point clouds were calculated, and the response curve of structural accuracy as a function of the scoring threshold was plotted. Statistical analysis revealed that when the scoring threshold τ... s Setting the threshold to 0.6 achieves an optimal balance between structural accuracy and redundancy removal rate, effectively removing low-quality points without significantly losing structural information. Therefore, in a preferred embodiment of the present invention, the structural scoring threshold is preferably set to 0.6, and this value is used in all subsequent structural point confidence removal processes.

[0219] For the three weighting coefficients α in the structural confidence scoring function θ α ρ With α ξThe method employed was a weighted training approach based on statistical contributions to structural stability. The specific steps included: first, dividing multiple tower and conductor structural regions into typical structural sample blocks; then, normalizing and comparing the response differences of each feature index (directional consistency, spatial density, and similarity score) between labeled and unlabeled structural regions; and calculating the relative information gain of each feature dimension on the structural recognition capability. Statistical analysis revealed that the structural similarity score index ξ... q The normal consistency index θ contributes the most to the determination of structural stability, accounting for approximately 50%. q Approximately 30% of the space is compact, according to the index ρ. q It accounts for approximately 20%. Therefore, it is preferable to set the three weighting coefficients as α. θ =0.3 α ξ =0.5.

[0220] Furthermore, step S500 sets the voxel mesh size and thinning weight for the tower point cloud and conductor point cloud after the deletion operation, and performs adaptive structural thinning processing, including the following steps E1-E3:

[0221] E1: Based on the gradient distribution of the change in the normal vector direction of each point in the tower point cloud in the vertical direction, construct a vertical structural trend estimation curve; divide the tower point cloud into multiple structural segments according to the curve; set a constant height interval as the voxel division unit size in each structural segment, and extract representative points covering the key nodes of the structural trend curve to form a representative tower point set.

[0222] E2: Based on the arrangement direction of each point in the traverse point cloud, generate the fitting path of the traverse principal axis; calculate the distance offset value of each point to the fitting path; use the fitting path as the reference axis to divide the projection segments with fixed spacing, and extract the points with the same arrangement direction and the minimum local curvature in each segment to form the representative point set of the traverse.

[0223] E3: Perform edge continuity analysis on the representative point set of towers and the representative point set of conductors respectively, and remove boundary points with abrupt changes in position distribution and redundant points with local density anomalies exceeding the preset redundancy threshold.

[0224] Furthermore, in step E1, firstly, for each point p in the tower point cloud... i Extract its unit normal vector n i , construct its neighborhood The neighborhood search radius is set to r = 0.25m. The angle between the normal vectors of all points in the neighborhood is statistically analyzed, and the angle at point p is calculated. i The magnitude of the change in the direction of the normal vector θ i Defined as:

[0225]

[0226] Where, n j Represents the neighborhood point p j The unit normal vector; θ represents the number of points in the neighborhood; i This indicates the structural gradient characteristics at that point; a larger value indicates a more significant abrupt change in the normal direction.

[0227] Then, all points are determined according to their height coordinates z. i The point cloud is divided into continuous height intervals, with an optimal height step size of Δz = 1.0m for each interval. The point cloud is then projected onto the vertical direction, and the normal variation amplitude θ within each height interval is statistically analyzed. i The average value is used to construct the height-gradient distribution function:

[0228]

[0229] Where N z This represents the number of points within the current interval. A vertical structural gradient trend curve is fitted using this function and used as the basis for structural segmentation. The tower structure is divided into multiple structural segments along the height direction, each segment having a different length. The dividing point is the location of a sharp gradient change, enhancing the expressive power of the segmented structure.

[0230] Within each structural segment, set the voxel division unit size V. d The preferred setting method is:

[0231] V d =max(δ z ,d avg )

[0232] Where, δ z Indicates the height range of the current paragraph; d avg This represents the average distance between points within the paragraph;

[0233] This setting ensures that the voxel size can cover the main structural morphology without excessively compressing the data density.

[0234] Next, key node extraction is performed on each structural segment. Equally spaced sampling planes are generated along the vertical direction, with an optimal interval of 0.5m. On each sampling plane, the point where the normal change within that layer is most stable (i.e., θ) is statistically analyzed. i The smallest point is selected, and its spatial center point is chosen as the representative point. Finally, the set of representative points extracted from all paragraphs constitutes the representative point set of the tower.

[0235] Furthermore, in step E2, firstly, for each point q in the traverse point cloud set... j Build neighborhood based on its spatial location The search radius is set to r = 0.4m. Principal direction fitting is performed within the neighborhood, i.e., principal component analysis (PCA) is executed to obtain the principal direction vector v for each point. j This indicates a local arrangement trend. By traversing all points, a sequence of points with the same continuous direction (i.e., the included angle is less than a set threshold, preferably 15°) is selected and merged to form a set of continuous conductor segments.

[0236] Next, the center point of each traverse segment is selected and connected according to their spatial order to form a preliminary sequence of principal axis points {Q1, Q2, ..., Q...}. n Perform B-spline fitting on the point sequence to generate the principal axis path function of the traverse in three-dimensional space:

[0237]

[0238] in, B represents the spatial representation function of the main axis path, where the parameter s is the normalized path length; i (s) represents the B-spline basis function; P i The coordinates of the control points are represented by k; k represents the spline order, preferably 3.

[0239] After fitting, for each point q in the original traverse point cloud j Calculate the minimum Euclidean distance from the fitted principal axis path, denoted as:

[0240]

[0241] The offset value d j This serves as an evaluation criterion for the consistency of the point and principal axis structure. To implement regional thinning, a traverse path function is used. Using the reference axis, the parameter space is divided into segments [s] according to a fixed path spacing Δs = 1.0m. i ,s i+1 ], which corresponds to a three-dimensional projection segment.

[0242] Within each projection segment, collect the set of traverse points falling within that segment. And perform the following two-stage screening:

[0243] Consistency filtering of arrangement direction: Calculate the local principal direction v of each point within the segment. j t(s) of the fitted path tangent direction i The included angle φ j , screen out φ j Anomalies >20°;

[0244] Minimum local curvature filter: For the remaining set of points, calculate the local curvature κ of the three neighborhoods of each point. j The point with the smallest curvature is selected as the representative point of this segment:

[0245]

[0246] Finally, the representative points extracted from each projection segment constitute the traverse representative point set. Used for subsequent edge removal and structure preservation thinning.

[0247] Furthermore, in step E3, the boundary point identification rule for abrupt changes in positional distribution is as follows: each representative point is sorted along the principal axis of its structure to construct an ordered spatial arrangement sequence. Based on this, the Euclidean distance difference between each point and its preceding and following adjacent points is calculated, and the mean and standard deviation of the entire difference sequence are calculated.

[0248] If the distance between a point and its neighbors deviates significantly from the normal mean, specifically if the difference in distance between adjacent points exceeds the overall mean plus twice the standard deviation, it is considered an abnormal jump in the local structure. Such points are identified as boundary points of abrupt changes in location distribution, typically appearing at the edge of the point cloud, in areas with missing data, or in areas with structural occlusion. They cannot effectively represent the continuous structural trend and must be removed.

[0249] In this invention, to ensure uniform and sparse point cloud and avoid computational burden caused by redundant data, local density analysis needs to be performed on each representative point. Specifically, a three-dimensional spherical neighborhood with a fixed radius (e.g., 0.5 meters) is constructed centered on each representative point, and the number of representative points within this neighborhood is counted, serving as the local density value for that point.

[0250] Subsequently, statistical analysis was performed on the density values ​​of all representative points, and the 90th decimal place was preferentially used as the basis for the density reference threshold. Based on this, an empirical correction coefficient (e.g., 1.1 times) was introduced to generate the final redundancy density threshold. When the density value of a point exceeds this threshold, it indicates that the point is in an over-dense region, possibly due to scan overlap or edge fusion causing dense accumulation. This point is considered redundant and needs to be removed.

[0251] Further, step S600 performs fusion registration processing on the thinned tower point cloud and the conductor point cloud, minimizing the structural consistency loss function. The minimized structural consistency loss function includes the tower principal axis offset term, the conductor curve residual term, and the spatial offset term of structural feature points. The fusion registration output result is used as the final point cloud data, including the following steps F1-F5:

[0252] F1: Extract a set of structural feature points from the tower point cloud; fit the tower principal axis vector in the vertical direction based on the set of structural feature points, and calibrate the coordinates of the tower principal axis starting point;

[0253] F2: Extract a continuously arranged set of points from the conductor point cloud; perform spatial fitting based on this set to generate a conductor curvature trend function;

[0254] F3: Based on the temporal or spatial overlap of point cloud acquisition, calibrate the correspondence between structural feature points of tower point clouds and conductor point clouds;

[0255] F4: Construct a structural consistency loss function, which includes: a tower principal axis offset term, representing the angular deviation between the tower point cloud principal axis vector and the reference principal axis vector, as well as the Euclidean distance residual between the principal axis starting points; a traverse curve residual term, representing the integral value of the curvature offset residual between the traverse point cloud curvature trend function and the reference curvature function in the corresponding segment; and a structural feature point spatial offset term, representing the sum of squared three-dimensional spatial coordinate residuals between corresponding structural feature points.

[0256] F5: Minimize the structural consistency loss function to solve the optimal spatial transformation matrix, apply the spatial transformation matrix to perform fusion registration operation on the tower point cloud and the conductor point cloud, and output the fused and registered point cloud data as the final point cloud data result.

[0257] Specifically, in step F1, a set of structural feature points is first extracted from the thinned tower point cloud data. These structural feature points are preferably points distributed along the vertical direction with significant inflection points of normal vector change, such as structural junctions like tower platforms, partitions, and crossarms. By analyzing the gradient curve of the normal vector change in the vertical direction, local extreme points are identified as candidate structural feature points. A spatial distribution screening strategy is then used to eliminate duplicate and dense points, retaining a discrete and uniformly distributed set of points.

[0258] After obtaining the set of structural feature points, principal component analysis (PCA) is used to fit the set to the principal axis, obtaining the principal axis direction vector. During the fitting process, the first principal direction of the covariance matrix is ​​set as the vertical principal axis of the tower, and the feature point with the lowest height in the point cloud is extracted as the starting point of the principal axis for subsequent spatial alignment with the reference tower model.

[0259] Specifically, in step F2, continuous point segments are extracted from the traverse point cloud. By setting a maximum allowable spacing threshold between adjacent points, point segments that are closely connected and aligned in the same direction are selected to form a candidate set of continuous curves. Subsequently, cubic B-spline curve fitting is performed on each continuous point segment to construct the curvature trend function of the traverse in space.

[0260] Specifically, in step F3, in order to achieve spatial fusion and registration of the tower and conductor point clouds, it is necessary to establish a correspondence between structural feature points between the two types of point clouds. In this embodiment, by combining the timestamp information during point cloud acquisition with spatial overlap region analysis, coarse matching is performed between the tower structural feature points and the conductor key nodes.

[0261] Specifically, the tower principal axis offset term is characterized by the angular error between the fitted principal axis direction of the tower point cloud and the preset reference principal axis direction, as well as the spatial position difference between the principal axis starting point and the reference principal axis starting point. The angular deviation reflects the vertical stability of the overall tower structure; if the angular deviation is too large, it will seriously affect the effectiveness of registration and the accuracy of structural matching. The Euclidean distance residual between the principal axis starting points characterizes the initial positioning error of the tower point cloud and the reference tower in spatial coordinates. The consistency of structural orientation and position is evaluated by combining the two.

[0262] The traverse curve residual term is obtained by curve fitting the traverse point cloud to obtain the spatial curvature trend function of the traverse. This function is then compared with the pre-defined reference traverse curvature trend function in corresponding spatial segments for piecewise integration error evaluation. Specifically, a numerical integration method is used to calculate the cumulative squared difference between the two curvature functions in each corresponding segment, thereby obtaining the consistency error of the curvature trend. This effectively reflects the morphological changes of the traverse structure before and after registration, ensuring that the registered traverse structure can accurately represent the actual spatial morphology.

[0263] The structural feature point spatial offset term selects the pairs of structural feature points matched in step F3 from the point clouds of towers and conductors, and calculates the sum of squares of the coordinate differences between each pair of feature points in three-dimensional space. This error can directly characterize the registration accuracy at the microscale, ensuring that corresponding points of local features can be accurately aligned and preventing local deformation or misalignment.

[0264] To effectively integrate the three different types of error terms, this embodiment further sets weight parameters, with the tower main axis offset term having the highest weight, followed by the conductor curvature trend residual term, and the structural feature point spatial offset term having the lowest weight. The optimization of the weights was determined through experimental analysis of multiple distribution network point cloud instance data. Specifically, the weight of the tower main axis offset term is 0.5, the weight of the conductor curvature trend term is 0.3, and the weight of the structural feature point spatial offset term is 0.2. This weight configuration ensures the priority of the overall structural orientation of the tower point cloud and the spatial curve shape of the conductor in the fusion registration process, while also taking into account the feature point matching accuracy.

[0265] After the structural consistency loss function is clearly defined, further optimization is performed to minimize the loss function. Specifically, a rigid spatial transformation model is used for registration, and a nonlinear optimization method is used to iteratively search for the optimal combination of spatial rotation matrix and translation vector until the loss function value reaches the preset convergence condition. During the optimization process, the initial coarse registration result is used as the starting point, and the termination condition is set as the loss function change being less than a set threshold or reaching the maximum number of iterations, thereby ensuring the convergence of the optimization process.

[0266] Finally, based on the obtained optimal spatial transformation matrix, a unified spatial transformation is performed on the tower point cloud and the conductor point cloud respectively to obtain the complete distribution network point cloud data results after fusion and registration.

[0267] In a preferred embodiment of the present invention, the fitting of the tower's main axis direction is based on the eigenvalue decomposition method of the covariance matrix, as shown in the formula:

[0268]

[0269] Where C represents the covariance matrix of the structural feature points of the tower point cloud; N represents the total number of structural feature points of the tower; x i Represents the spatial coordinates of the i-th structural feature point of the tower; The value represents the mean of the spatial coordinates of the set of characteristic points of the tower structure; T represents the matrix transpose operator.

[0270] The specific mathematical expression for performing cubic B-spline fitting on the guide point cloud is as follows:

[0271]

[0272] Where C(s) represents the curve fitting function for the conductor point cloud; s represents the normalization parameter, indicating the position along the curve length; P i B represents the coordinates of the i-th control point; i,k (s) represents the i-th B-spline basis function (order k = 3); m+1 represents the total number of control points used in the B-spline fitting curve.

[0273] The overall expression for the structural consistency loss function is:

[0274]

[0275] The explanations for each part are as follows:

[0276] Tower spindle offset term:

[0277] θ a =arccos(a′·a),d a =||o′-o||

[0278] Where a′ represents the principal axis direction vector of the registered tower; a represents the principal axis direction vector of the reference tower; θ a Indicates the spindle included angle deviation; o′ represents the coordinates of the starting point of the spindle after registration; o represents the coordinates of the starting point of the spindle of the reference tower; d a This represents the Euclidean distance between the starting points of the spindle.

[0279] Conductor curvature residual term:

[0280]

[0281] Where C′(s) represents the curvature trend function after fitting the traverse point cloud; C(s) represents the curvature trend function of the reference traverse; s1, s2 represent the start and end parameters of the integral calculation segment of the fitted curve; E c The integral residual error value represents the trend of conductor curvature.

[0282] Structural feature point spatial offset term:

[0283]

[0284] Where, p i q represents the spatial coordinates of the i-th registered structural feature point; i E represents the spatial coordinates of the i-th reference feature point; p This represents the sum of squared spatial coordinate residuals between all pairs of structural feature points; N represents the total number of pairs of structural feature points.

[0285] It should be noted that λ1 = 0.5: the weighting coefficient of the tower main axis offset term; λ2 = 0.3: the weighting coefficient of the conductor curvature residual term; λ3 = 0.2: the weighting coefficient of the feature point spatial offset term; the three coefficients satisfy the condition: λ1 + λ2 + λ3 = 1.

[0286] The optimal rigid spatial transformation matrix T is obtained by minimizing the loss function using an optimization algorithm. The spatial transformation is defined as follows:

[0287] x′=T(x=Rx+t

[0288] Where x represents the spatial coordinates of the tower or conductor point cloud to be transformed; x′ represents the transformed spatial coordinates; R represents the 3×3 rotation matrix to be optimized; and t represents the translation vector to be optimized.

[0289] The rotation matrix R and translation vector t are solved by optimization methods (such as the Levenberg-Marquardt algorithm) to minimize the structural consistency loss function L.

[0290]

[0291] Among them, T * Let L(R,t) represent the optimal spatial transformation matrix; argmin represents the combination of parameters that minimizes the function; L(R,t) represents the loss function L as a function of the rotation and translation parameters.

[0292] Finally, using T * Spatial transformations are performed on the point clouds of towers and conductors respectively to achieve fusion registration.

[0293] Example 3 is an embodiment of the present invention, which provides a noise reduction and thinning system for point cloud data of power distribution network, including: a construction module configured to construct tower structure models and conductor structure models;

[0294] The feature extraction and classification module is configured to extract the spatial height, normal vector direction, neighborhood point arrangement direction and local curvature of each point in the three-dimensional point cloud data of the power distribution network, and perform similarity calculations with the tower structure model and conductor structure model respectively, and classify the point cloud into tower point cloud, conductor point cloud and background point cloud based on the similarity score results.

[0295] The structure scoring module is configured to perform structure confidence scoring on each point in the classification results, extract the normal vector direction, neighborhood geometry and similarity score, and generate a structure score value.

[0296] The removal module is configured to delete tower point cloud points and conductor point cloud points whose structural score values ​​are lower than a preset threshold, and to delete all background point cloud points;

[0297] The thinning module is configured to set the voxel mesh size and thinning strategy for the removed tower point cloud and conductor point cloud respectively, and perform structure-adaptive thinning processing.

[0298] The registration module is configured to perform a fusion registration operation on the thinned tower point cloud and the conductor point cloud, minimizing the structural consistency loss function.

[0299] This embodiment also provides an electronic device applicable to a method for noise reduction and thinning of distribution network point cloud data, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the method for noise reduction and thinning of distribution network point cloud data as proposed in the above embodiment.

[0300] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements a method for noise reduction and thinning of distribution network point cloud data as proposed in the above embodiment.

[0301] The storage medium proposed in this embodiment and the method for noise reduction and thinning of point cloud data in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0302] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0303] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for noise reduction and thinning of point cloud data in a power distribution network, characterized in that: The method comprises the following steps: Constructing a tower structure model and a conductor structure model; the tower structure model comprises a height range, a vertical direction feature, and a local normal concentration threshold; the conductor structure model comprises a curvature range, a linear extension direction, and a continuous point spacing threshold; For each point in the three-dimensional point cloud data of the power distribution network, the spatial height, the normal vector direction, the neighborhood point arrangement direction, and the local curvature are extracted, and similarity calculation is performed on the tower structure model and the conductor structure model respectively, and the point cloud is classified into a tower point cloud, a conductor point cloud, and a background point cloud according to the similarity score results; For each point in the classification result, structure confidence score processing is performed, the input features include the normal vector, the neighborhood geometric shape, and the similarity score, and a structure score value is output; The tower point cloud points and the conductor point cloud points with a score value lower than a preset threshold are deleted, and all the background point cloud points are deleted; After the deletion operation, the tower point cloud and the conductor point cloud are set with a voxel grid size and a thinning weight, and structure adaptive thinning processing is performed; After the thinning processing, the tower point cloud and the conductor point cloud are subjected to fusion registration processing, a structure consistency loss function is minimized, the structure consistency loss function comprises a tower main shaft offset term, a conductor curve residual term, and a structure feature point space offset term, and the fusion registration output result is used as the final point cloud data.

2. The power distribution point cloud data denoising and sparsifying method of claim 1, wherein: The construction of the tower structure model and the conductor structure model comprises the following steps: The spatial height feature of each point in the point cloud is extracted, the average height value of a local region is calculated, the points belonging to a set height range are screened, and a first candidate point set is formed; The vertical direction feature is extracted in the first candidate point set, the angle relationship between the main direction of the point and the vertical direction is calculated, the points meeting the vertical direction feature requirement are identified, and a second candidate point set is formed; The neighborhood of each point in the second candidate point set is subjected to normal vector statistics, a normal vector consistency index is calculated, the points meeting the local normal concentration threshold are screened, and a tower structure point set is constituted; 3. The power distribution point cloud data denoising and sparsifying method of claim 2, wherein: Based on the statistical features of the tower structure point set, the tower structure model comprising the height range, the vertical direction feature, and the local normal concentration threshold is defined. The construction of the tower structure model and the conductor structure model further comprises the following steps: In the point cloud data, a continuous point sequence with a distance between adjacent points less than a set threshold is identified, a point segment meeting the continuous point spacing threshold is screened, and a conductor candidate point set is formed; The local curvature feature of each point sequence in the conductor candidate point set is calculated, the point segments with a curvature exceeding a set curvature range are excluded, and the point segments with a curvature meeting the requirement are retained; In the retained point segments, direction vector fitting processing is performed, the linear extension direction is extracted, the point segments with a direction change degree within an allowable range are identified, and a conductor structure point set is formed; Based on the curvature range, the linear extension direction, and the point spacing parameters of the conductor structure point set, the conductor structure model is constructed.

4. The power distribution point cloud data denoising and sparsifying method of claim 3, wherein: The point cloud is classified into a tower point cloud, a conductor point cloud, and a background point cloud according to the similarity score result, including sequentially setting a weighted priority factor in tower structure model scoring according to the distance between the spatial height of each point and the estimated center of the tower main shaft; calculating the difference between the spatial height, the normal vector direction, and the neighborhood normal concentration of each point and the height range, the vertical direction feature, and the neighborhood normal concentration threshold set in the tower structure model, and combining the weighted priority factor to form a tower structure model score value; Combination calculation is performed on the direction deviation between the arrangement direction of the neighborhood points of each point and the linear extension direction set in the conductor structure model, the difference between the local curvature and the curvature range set in the conductor structure model, and the deviation between the distance between adjacent points and the continuous point spacing threshold set in the conductor structure model to form a conductor structure model score value; The average of the tower structure model score values of the points in the neighborhood range centered on each point and the average of the conductor structure model score values are extracted, and the tower structure model score value and the conductor structure model score value of each point are weighted and corrected to obtain a corrected tower structure model score value and a corrected conductor structure model score value; When the corrected tower structure model score value is greater than the corrected conductor structure model score value, the score difference between them is greater than the score difference threshold, and the corrected tower structure model score value is greater than the classification confidence threshold, it is determined as a tower point cloud category; When the corrected conductor structure model score value is greater than the corrected tower structure model score value, the score difference between them is greater than the score difference threshold, and the corrected conductor structure model score value is greater than the classification confidence threshold, it is determined as a conductor point cloud category; In the case of not meeting the tower point cloud category and the conductor point cloud category, it is determined as a background point cloud category.

5. The power distribution point cloud data denoising and sparsifying method of claim 4, wherein: The input features include a normal vector, a neighborhood geometry, and a similarity score, and the output structure score value includes extracting the normal vector direction change feature of each point, the local distribution geometry feature of the points in the neighborhood range, and the similarity intensity feature corresponding to the category structure model score value, and constructing a structure confidence feature vector; The normal vector direction change amplitude, the neighborhood point spatial compactness index, and the similarity score intensity index are respectively converted into confidence sub-feature values; According to a preset weighted fusion rule, the confidence sub-feature values are combined to construct a structure confidence score function; The structure confidence feature vector is input into the structure confidence score function to generate a structure score value of each point; When the structure score value of a certain tower point cloud point or conductor point cloud point is lower than a structure score threshold, the point is deleted from the corresponding point cloud set; All points in the background point cloud set are deleted.

6. The power distribution point cloud data denoising and sparsifying method of claim 5, wherein: The execution of the structure adaptive thinning processing includes constructing a vertical direction structure trend estimation curve based on the gradient distribution of the normal vector direction change amplitude of each point in the tower point cloud; The tower point cloud is divided into multiple structure paragraphs according to the curve; a constant height interval is set as a voxel division unit size in each structure paragraph, and a representative point set of the tower is formed by extracting the representative points covering the key nodes of the structure trend curve; Based on the arrangement direction of each point in the conductor point cloud, a conductor main shaft fitting path is generated; the distance offset value of each point to the fitting path is calculated; Taking the fitting path as the reference axis, the projection section with fixed interval is divided, and the points with consistent arrangement direction and minimum local curvature in each section are extracted to form the conductor representative point set; The edge continuity analysis is performed on the tower representative point set and the conductor representative point set respectively, and the boundary points with position distribution mutation and the redundant points with local density higher than the preset redundancy threshold are removed.

7. The power distribution point cloud data denoising and sparsifying method of claim 6, wherein: The structure consistency loss function includes, The structure feature point set is extracted from the tower point cloud; The tower main shaft vector is fitted based on the structure feature point set in the vertical direction, and the tower main shaft starting point coordinate is calibrated; The continuous arrangement point set is extracted from the conductor point cloud; the conductor curvature trend function is generated based on the set; The structure feature point corresponding pair between the tower point cloud and the conductor point cloud is calibrated based on the point cloud collection time or the spatial overlap relationship; The structure consistency loss function is constructed, which includes the tower main shaft offset term, the conductor curve residual term, and the structure feature point space offset term. The optimal spatial transformation matrix is calculated by minimizing and optimizing the structure consistency loss function, and the spatial transformation matrix is applied to perform the fusion registration operation on the tower point cloud and the conductor point cloud, and the fusion registered point cloud data is output as the final point cloud data result. It includes: The construction module is configured to construct the tower structure model and the conductor structure model; 8. A power distribution grid point cloud data denoising and sparsifying system, applying a power distribution grid point cloud data denoising and sparsifying method according to any one of claims 1-7, characterized in that, The feature extraction and classification module is configured to extract the spatial height, normal vector direction, neighborhood point arrangement direction and local curvature of each point in the power distribution network three-dimensional point cloud data, and perform similarity calculation with the tower structure model and the conductor structure model respectively, and classify the point cloud into tower point cloud, conductor point cloud and background point cloud according to the similarity score result; The structure scoring module is configured to perform structure confidence scoring processing on each point in the classification result, extract the normal vector direction, neighborhood geometry and similarity score, and generate the structure score value; The removal module is configured to delete the tower point cloud points and conductor point cloud points with structure score value lower than the preset threshold, and delete all background point cloud points; The thinning module is configured to set the voxel grid size and thinning strategy for the removed tower point cloud and conductor point cloud respectively, and perform structure adaptive thinning processing; The registration module is configured to perform fusion registration operation on the tower point cloud and the conductor point cloud after thinning processing, and minimize the structure consistency loss function. The processor executes the computer program to realize the steps of the power distribution network point cloud data denoising and thinning method in any one of claims 1 to 7. ​ 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. ​ 10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by a processor to implement the steps of the power distribution network point cloud data denoising and thinning method in any one of claims 1 to 7.

Images